The current challenge boils down to the tradeoffs of local LLM configuration. In other words, it comes down to implementing the best load balancing between the CPU and GPU for MoE architecture based LLMs. The next step will inevitably be improving the performance of the next layer. LM Studio introduced a load control switch in llama.cpp in version 0.3.23, which allows for tweaking. So, what about Ollama? As more and more examples emerge, the finer details will be ironed out, such as the need for the dropout control in GPT-2. 8GB of memory is now the norm even for smartphones. With i9-level performance, even expensive GPUs aren’t necessary, as they can be supplemented with NPUs or other “AI-specific” features. A community model will be essential and effective for diverse personal use.